Self-discharge diagnosis algorithm for lithium-ion battery packs based on remaining rechargeable capacity prediction for multi-stage dynamic charging scenarios

Abstract

This invention discloses a method for diagnosing self-discharge anomalies in energy storage battery packs based on short-term charging data, belonging to the field of battery safety monitoring technology. Addressing the shortcomings of existing methods that rely on complete charge-discharge cycles and are difficult to adapt to multi-stage varying operating conditions, this invention extracts features from multi-stage charging data, combines an autoencoder to achieve adaptive extraction of features under multiple operating conditions, and constructs a data-driven model fused with GCN-BiLSTM (Graph Convolutional Network-Bidirectional Long Short-Term Memory) to accurately estimate the RCC (Remaining Charging Capacity) of individual battery cells. Finally, based on the size, distribution, and changes of the RCC of each individual cell in the battery pack, we can diagnose overall battery pack inconsistencies, individual cell SOC inconsistencies, and self-discharge faults. This provides a guarantee for the safe operation of energy storage systems.

Description

Technical Field

[0001] This invention belongs to the field of intelligent management technology for energy storage batteries, and specifically studies a diagnostic method for abnormal self-discharge of individual cells in a battery pack containing more than 400 large-capacity cells under multi-stage charging conditions. Background Technology

[0002] Lithium-ion batteries have been widely used in energy storage due to their superior performance, including high energy density, high power density, long cycle life, and low self-discharge rate. However, because the voltage and power of a single lithium-ion battery are insufficient to meet the demands of energy storage systems, it is usually necessary to construct battery packs by connecting hundreds of individual batteries in series or parallel to meet the energy storage and output requirements of large-scale energy storage systems. Achieving efficient energy storage and conversion hinges on the consistency between individual lithium-ion batteries, which is also a crucial foundation for building high-performance energy storage battery packs. Self-discharge is a key indicator in evaluating the differences in the electrical performance of individual lithium-ion batteries. Generally, batteries with higher self-discharge rates exhibit faster capacity decay, poorer cycle characteristics, and shorter lifespan, thus exacerbating the inconsistency between individual batteries. This inconsistency not only significantly reduces the overall performance of the energy storage battery pack but may also increase the risk of system failure during operation. Currently, commonly used methods for detecting self-discharge in lithium-ion batteries mainly include the open-circuit voltage method, the capacity decay method, model-based methods, and data-driven methods. The open-circuit voltage method determines the self-discharge rate by monitoring the voltage change of the battery in an open-circuit state, but the test time is long and it is difficult to detect minute self-discharges. The capacity decay method assesses self-discharge by recording the capacity loss of the battery over a certain period of time, but its operation is complex and the test cycle is long. Model-based methods use equivalent circuit model parameters to fit the self-discharge behavior. Although the test efficiency is high, the requirements for model accuracy and parameter identification are high, and errors may exist in practical applications. These methods cannot simultaneously meet the requirements of high efficiency, accuracy, and adaptability. In recent years, artificial intelligence technology has developed rapidly and has been effectively applied in the field of battery safety early warning and fault diagnosis. Various data-driven methods are based on the historical operating data of the battery, without needing to explore its complex fault mechanisms. The basic steps of data-driven modeling include extracting domain battery self-discharge-related data and carefully selecting algorithms suitable for specific applications. For data-driven methods, the key is to develop advanced algorithms and feature extraction methods for specific applications. This invention proposes a data-driven approach for locating and diagnosing abnormal self-discharge cells in in-service energy storage battery packs based on short-term charging data. This method can quantify the evolution trend of self-discharge and distinguish between SOC (State of Charge) inconsistency and self-discharge inconsistency. Summary of the Invention

[0003] With the significant increase in demand for energy storage systems, new lithium-ion battery energy storage systems have seen tremendous development due to the superior performance of lithium-ion batteries. The performance of the battery pack is crucial to the overall operating efficiency and reliability of the energy storage system. Inconsistencies in lithium-ion battery packs can arise due to manufacturing processes and usage, and if not diagnosed and addressed promptly, can have serious consequences. In particular, self-discharge in lithium-ion batteries is often caused by internal short circuits resulting from metal impurities or burrs piercing the separator, posing a risk of thermal runaway. Therefore, timely and accurate location and diagnosis of abnormal cells has become a key indicator for evaluating the economic value and operational lifespan of energy storage systems, and has significant practical implications. However, since self-discharge is an inherent characteristic of electrochemical systems, all batteries experience some degree of self-discharge during normal use, making it difficult to accurately obtain the changing trend of individual cell self-discharge levels. Furthermore, existing evaluation methods typically rely on lengthy charge-discharge data or fixed feature extraction methods, which face significant limitations in real-world energy storage scenarios with multi-stage and varying operating conditions. This invention proposes a self-discharge anomaly diagnosis method for energy storage battery packs based on short-term charging data. By extracting multi-stage charging variation data and combining it with an autoencoder to adaptively extract multi-condition features, a data-driven model of GCN-BiLSTM (Graph Convolutional Network - Bidirectional Long Short-Term Memory) is constructed to accurately estimate the RCC (Remaining Charging Capacity) of individual battery cells. This effectively solves the problem that traditional methods rely on complete charge-discharge cycles and are difficult to adapt to complex charge-discharge strategies. This technology uses the changing trend of RCC value to monitor the degree of battery self-discharge in real time. Combined with sliding window statistics, it can distinguish between capacity decay and self-discharge anomalies caused by internal short circuits, realizing early risk warning (such as triggering an alarm when the weekly RCC decrease exceeds 5%). At the same time, it is compatible with multi-stage charging scenarios, providing a high-precision, low-latency safety diagnosis solution for energy storage systems. It has significant application value in fields such as cascade utilization screening, dynamic equalization optimization, and thermal runaway prevention.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A fault diagnosis algorithm based on historical operating data and utilizing deep learning to evaluate the Regenerative Computation (RCC) of batteries and thus detect abnormal self-discharge cells in the battery pack. It specifically includes the following six parts:

[0006] S1: During the charging and operation of the energy storage system, battery data is collected in real time and preprocessed through the BMS (Battery Management System). The specific process includes:

[0007] S11: Compared to the discharge conditions of battery packs in energy storage systems, the charging process is easier to control. The energy storage system's BMS collects charging data from the battery pack, including time, total voltage, current, maximum cell voltage, minimum cell voltage, maximum cell temperature, minimum cell temperature, and individual cell voltage. This data is then recorded and stored.

[0008] S12: Before fault diagnosis, the battery cluster data needs to be cleaned to improve data quality and increase the accuracy and reliability of the algorithm. Preprocessing of the data includes: deduplication, sorting out-of-order values, outlier removal, and missing value imputation. The selected valid battery data is then stored and recorded for later use.

[0009] S2: Constructing samples for subsequent model training. The specific process includes:

[0010] S21: The current switching point contains rich inconsistencies. Therefore, the input can be selected as voltage and current data at the current switching point. The system filters individual cells in the battery pack that have reached the upper cutoff voltage, defines current switching event detection conditions, and triggers data truncation when the charging current change rate exceeds a preset threshold ΔI_threshold. It also filters individual cells in the battery pack that have reached the upper cutoff voltage, and symmetrically extracts the voltage-current-temperature three-dimensional sequence of n sampling points centered on the switching point, mainly including the individual cell voltage sequence. V =[ V 1 , V 2 , ..., V n ], current sequence I =[ I 1 , I 2 , ..., I n and temperature sequence T =[ T 1 , T 2 , ..., T n The final construction of the sample can be represented as:

[0011]

[0012] S22: During battery pack charging, individual cells that have reached the upper cutoff voltage are considered nearly fully charged. At this point, their remaining charging capacity can be calculated by integrating the charging data after the sample termination point in amperes. The specific steps are as follows:

[0013] First, determine the time range. For each individual cell that reaches the upper cutoff voltage, define the time range […]. T 0, T 1]; Start time T 0: The start time of the remaining charging segment of a single battery cell, determined by the end point of the sample selected in S21; the end time... T 1: The actual end time of charging of a single battery cell, i.e., the time when the charging current drops to the termination threshold or charging stops after the constant current to constant voltage transition; then extract the charging data within the time range [ T 0, T Within [1], extract the current data sequence of the single cell { I ( t 0), I ( t 1), …, I ( t b )},in b This indicates the number of sampling points in the remaining charging segment; subsequently, the remaining charging capacity is calculated using the Ampere integral formula to determine the remaining charging capacity of each individual cell. Q :

[0014]

[0015] Finally, the remaining charging capacity of each individual cell that has reached the upper cutoff voltage is calculated to form a matrix of the remaining charging capacity of each cell: Q =[ Q 1 ，Q 2 ， … ，Q N ],in N This refers to the number of battery cells.

[0016] S23: Normalize the obtained samples. Use maximum-min normalization for voltage, and divide by the maximum value for current and remaining charging capacity to facilitate subsequent model training.

[0017] S3: Construct an encoder-decoder network. The input is the three-channel time sequence of charging data samples: voltage U, current I, and temperature T. The encoder extracts features through three one-dimensional convolutional layers and compresses them into the latent space. The output dimension is... p * q The intermediate feature vector ( p For the preset feature dimensions, q (Preset feature length); the decoder reconstructs the input data through deconvolution layers, and the loss function is defined as the sum of the three-channel mean square errors of the input and the reconstructed output:

[0018]

[0019] in, x k This refers to the input sample. x k recon This refers to reconstructing samples; intermediate features are mapped to standardized dimensions through a fully connected layer. p*q The feature vectors are used as input features for the capacity prediction model below.

[0020] S4: Construct a data-driven model using GCN-BiLSTM to estimate the remaining charge capacity of individual battery cells. Specifically, this includes:

[0021] S41: In the data construction process, the intermediate features are first divided into equal parts of a fixed length. w Each time segment is used to calculate the mean value of a node feature vector, and these nodes together form the vertex set of the graph. Then, the cosine similarity between the features of the nodes is calculated to establish the connection relationship of the edges. When the similarity between two nodes exceeds a preset threshold, a directed edge is established between the two nodes. All node pairs that meet the conditions will be connected by edges, and finally a directed graph structure based on the similarity of time series data is formed. The nodes retain the local statistical features of the original time series data, while the edges reflect the correlation strength between different time segments.

[0022] S42: This model employs a hybrid architecture of Graph Convolutional Neural Network (GCN) and Bidirectional Long Short-Term Memory (BiLSTM). During construction, a two-layer Graph Convolutional Network (GCNConv) is first used to extract features and perform message passing on the input graph data. Each GCN layer is followed by a ReLU activation function to increase non-linear expressive power. Then, graph-level feature pooling is performed using global_mean_pool to aggregate the features of all nodes into a unified graph representation. Next, the pooled features are input into a BiLSTM layer for temporal modeling to capture long-term dependencies in the sequence. Finally, a fully connected layer maps the hidden states of the BiLSTM to the prediction target space, thus completing the end-to-end mapping process from graph structure features to the final predicted value. The entire network model is trained to optimize the loss function. MAE (Mean Absolute Error) is chosen as the loss function to quantify the difference between the predicted and true values. The formula is as follows:

[0023]

[0024] in, y i For the true value, yi p These are the model's predicted values. m This represents the number of samples.

[0025] S43: During model training, the entire training process was built on the PyTorch framework. The model structure adopted a hybrid architecture of GCN and BiLSTM. The GCN part contained two layers of graph convolutional networks, each followed by a ReLU activation function, while the BiLSTM layers were used for sequence modeling. During training, the Adam optimizer was used for parameter optimization, with an initial learning rate set to 0.005 and a total of 1000 training epochs. The model training adopted a batch processing approach, with a batch size set to 512. Data was loaded in batches and randomly shuffled using DataLoader. In each training epoch, the model sequentially performed forward propagation to calculate the predicted values, used L1Loss to calculate the loss, and then updated the model parameters through backpropagation.

[0026] S5: Based on the trained model, calculate the remaining rechargeable charge (RCC) for each unit in each cycle:

[0027] S51: Following the method in S21, input samples are extracted from the charging data of each charging cycle, and then the samples are input into the trained model to estimate the value of each individual in the first charging cycle. j The remaining charging capacity after one charge cycle to the upper cutoff voltage. Q j =[ Q 1j ，Q 2j ，…，Q ij ，Q Nj ],in N Indicates that the battery pack has a total of N Individual battery cells, Q ij Indicates the first i The monomer in the first j The remaining charging capacity after reaching the upper cutoff voltage in one cycle.

[0028] S52: Due to different battery charging strategies, the current switching points are not entirely the same. This means the starting point for calculating the remaining rechargeable capacity is not consistent. To ensure a relatively stable RCC value in each calculation, we need to... Q j Normalization is performed to obtain the RCC values ​​of cells that cannot reach the upper cutoff voltage, thereby eliminating system bias and improving the stability of the calculation results. The specific calculation formula is as follows:

[0029] in, RCCij In the j-th loop, the first... i The remaining rechargeable capacity of each cell; min ( Q j ) is the first j The minimum remaining charge capacity of all cells in each cycle. The estimated RCC of all cells in the j-th cycle is obtained. RCC j = [ RCC 1j ， RCC 2j ,..., RCC Nj ]:

[0030]

[0031] S5: The RCC (Recovery Capacity) of a lithium battery can effectively reflect the inconsistency between individual battery cells. Generally, cells with smaller RCCs have higher SOCs and smaller capacities, while cells with larger RCCs have lower SOCs and larger capacities. Based on the size, distribution, and variation of the RCC of each cell in the battery pack, we can diagnose overall battery pack inconsistencies, individual cell SOC inconsistencies, and self-discharge faults. The specific process includes:

[0032] S51: First, the overall inconsistency of the battery pack is measured. To quantify the overall inconsistency of the battery pack, the range of RCC values ​​is used as the metric. The range is defined as:

[0033]

[0034] Where: max( RCC j ): No. j The maximum value of the individual RCC in each cycle; min( RCC j ): No. j The minimum RCC of a single unit in each cycle. Range RCC This range directly reflects the range of RCC differences among individual cells within the battery pack and can be used to characterize the consistency of the battery pack. The larger the range, the more severe the inconsistency among the individual cells within the battery pack.

[0035] S52: Next is the inconsistency of SOC in individual cells, which can be calculated by subtracting the difference between the RCC value of a single cell and the median RCC. The median is chosen for its noise resistance and accurate reflection of the central tendency of the data. The formula is as follows:

[0036]

[0037] in:D ij : No. j In the loop, the first... i The difference between the RCC of each individual and the median; Median ( RCC j ): No. j The median of the RCC set in each cycle. Individual difference. D ij It characterizes the deviation of its RCC from the concentration level of the battery pack, thus reflecting the inconsistency of the SOC of the cell.

[0038] S53: When a cell in a battery pack experiences a self-discharge fault, its RCC value will show an abnormal decrease, and the slope of its RCC value will differ significantly from that of a normal cell. Based on this principle, we can use the sliding window local consistency analysis method to calculate the RCC slope of each cell in the battery pack, thereby identifying the self-discharging cell. The sliding window local consistency analysis method is a highly efficient anomaly detection method, especially suitable for time series data with high noise or large fluctuations. Through the sliding window local consistency analysis method, we can further diagnose battery self-discharge faults.

[0039] First, define the sliding window: based on the sampling frequency of the battery pack cycling data and the analysis requirements, set a reasonable window length. W The window size can be set to 5 or 10 cycles, depending on the volatility of the data. An appropriate sliding step size should also be selected. Δ (e.g., 1 or 2) Ensure the window covers all data points. The range of the sliding window is... T k =[ j , j + W -1], the formula for calculating the rate of change of RCC is:

[0040]

[0041] △ RCC i,k It is a matrix representing the RCC of each battery cell in the window. T k Rate of change within.

[0042] Calculation within the time window T k The local mean within the time window is used to calculate the local deviation of each individual within this time window, thereby measuring the consistency of change of each individual within this time window. The specific formula is as follows:

[0043]

[0044] Where: indicates within the time window T K The local mean of each time window within the period, while S i,k It is the local deviation, which represents the difference between the individual change and the local average change.

[0045] S54: Finally, the RCC range obtained for each cycle can be calculated based on historical data. RCC The difference between the RCC of a single monomer and the median D ij Use the 95th percentile to set the threshold. thrshold1 , thrshold2 This is used to assess the overall inconsistency of the battery pack and the inconsistency of the state of charge (SOC) of individual battery cells. It also involves calculating local deviations based on historical data. S i,k 95th percentile threshold setting thrshold3 and multiple consecutive time windows exceeding the threshold setting thrshold4 To locate and diagnose the faults of self-discharging cells.

[0046] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0047] This invention proposes a method for locating and detecting self-discharge cells in large-scale energy storage battery clusters containing more than 400 cells.

[0048] This invention utilizes short-term charging data to generate accurate tag data to support deep learning models, overcoming the limitations of traditional methods that rely on long-term, complete cyclic data, and significantly reducing data acquisition time and computational costs. Simultaneously, this method greatly expands the model's applicable operating conditions, including different charge / discharge rates and complex operating environments, ensuring its robustness and universality in real-world scenarios. By combining rapid tag generation with deep learning models, rapid diagnosis of lithium battery self-discharge faults is achieved, effectively improving the operating efficiency of energy storage systems.

[0049] This invention improves the accuracy of RCC estimation and the reliability of diagnosis by constructing RCC tags based on real-world data, overcoming the shortcomings of traditional methods that rely on simulation data or ideal conditions. Simultaneously, it employs a sliding window technique to generate multiple samples within a single cycle, extracting more useful information from local features of the time series, thus enhancing data utilization efficiency and model accuracy. The introduction of the sliding window method not only improves the ability to identify noise and anomalies but also expands the model's adaptability to dynamic operating conditions, making it more sensitive and reliable in practical applications.

[0050] This invention constructs a multi-condition adaptive feature extraction model with an encoder-decoder architecture. By reconstructing the input, it maps input data under different conditions to the same latent space and extracts intermediate features, thus adapting to changes in operating conditions.

[0051] This invention constructs a graph-based time series graph from battery cell time series data through segmentation and cosine similarity. It combines graph convolutional networks (GCN) to extract graph features and bidirectional long short-term memory networks (BiLSTM) to capture temporal dependencies, forming an end-to-end prediction framework that avoids tedious manual feature design. At the same time, it improves the accuracy and efficiency of the model through global feature pooling and optimized training strategies, demonstrating a powerful ability to model complex temporal relationships and graph data.

[0052] This invention constructs a hierarchical diagnostic framework using the RCC index, systematically quantifying and diagnosing battery status from overall pack consistency to individual cell SOC inconsistency and self-discharge faults. This includes introducing the RCC range to measure overall consistency, using deviation from the median to assess individual cell SOC inconsistency, and detecting self-discharge faults through sliding window local consistency analysis, enhancing the precision and robustness of the diagnosis. Simultaneously, dynamically setting multi-level thresholds based on historical data effectively improves the method's adaptability and noise resistance, making it suitable for health management and predictive maintenance of complex battery packs. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort:

[0054] Figure 1 A flowchart for detecting abnormal self-discharge cells in a battery cluster;

[0055] Figure 2 A flowchart of the algorithm for estimating the remaining charging capacity;

[0056] Figure 3 A framework diagram for using RCC for hierarchical diagnosis. Detailed Implementation

[0057] To achieve the above objectives, the present invention provides the following technical solution:

[0058] A fault diagnosis algorithm that uses deep learning to evaluate the Regenerative Computation (RCC) of a battery based on historical operating data, thereby detecting abnormal self-discharge cells in the battery pack. For example... Figure 1 As shown, it specifically includes the following 6 steps.

[0059] Step 1: During the charging and operation of the energy storage system, the BMS (Battery Management System) collects battery data in real time and performs data preprocessing.

[0060] Compared to the discharge conditions of battery packs in energy storage systems, the charging process is easier to control. The energy storage system's BMS collects charging data from the battery pack, including time, total voltage, current, maximum and minimum cell voltage, maximum and minimum cell temperature, and individual cell voltage. This data is recorded and stored. Before fault diagnosis, the battery cluster data needs to be cleaned to improve data quality and increase the accuracy and reliability of the algorithm. Preprocessing of the data includes: deduplication, sorting, outlier removal, and missing value imputation. The selected valid battery data is then stored and recorded for later use.

[0061] Step 2: Construct samples for subsequent model training. The specific process includes:

[0062] First, the current switching point contains rich inconsistencies. Therefore, the input can be selected as voltage and current data at the current switching point. The system filters individual cells in the battery pack that have reached the upper cutoff voltage, defines current switching event detection conditions, and triggers data extraction when the charging current change rate exceeds a preset threshold ΔI_threshold. It then filters individual cells in the battery pack that have reached the upper cutoff voltage, and symmetrically extracts the voltage-current-temperature three-dimensional sequence of n sampling points centered on the switching point, mainly including the individual cell voltage sequence. V =[ V 1 , V 2 , ..., V n ], current sequence I =[ I 1 , I 2 , ..., I n and temperature sequence T =[ T 1 , T 2 , ..., T n The final construction of the sample can be represented as:

[0063]

[0064] Then, the output samples are constructed. During the battery pack charging process, individual cells that have reached the upper cutoff voltage are considered to be nearly fully charged. At this point, their remaining charging capacity can be calculated by integrating the charging data after the sample termination point in amperes. The specific steps are as follows:

[0065] First, determine the time range. For each individual cell that reaches the upper cutoff voltage, define the time range […]. T 0, T 1]; Start time T 0: The start time of the remaining charging segment of a single battery cell, determined by the end point of the sample selected in S21; the end time... T 1: The actual end time of charging of a single battery cell, i.e., the time when the charging current drops to the termination threshold or charging stops after the constant current to constant voltage transition; then extract the charging data within the time range [ T 0, T Within [1], extract the current data sequence of the single cell { I ( t 0), I ( t 1), …, I ( t b )},in b This indicates the number of sampling points in the remaining charging segment; subsequently, the remaining charging capacity is calculated using the Ampere integral formula to determine the remaining charging capacity of each individual cell. Q :

[0066]

[0067] Finally, the remaining charging capacity of each individual cell that has reached the upper cutoff voltage is calculated to form the remaining charging capacity: Q =[ Q 1 ，Q 2 ， … ，Q N ],in N This represents the number of battery cells. The obtained samples are then normalized. Maximum-minimum normalization is used for voltage, while current and remaining charging capacity are normalized by dividing by the maximum value, facilitating subsequent model training.

[0068] Step 3: As Figure 2 As shown, an encoder-decoder network is constructed. The input is a three-channel time sequence of charging data samples: voltage U, current I, and temperature T. The encoder extracts features through three one-dimensional convolutional layers and compresses them into the latent space. The output dimension is... p * q The intermediate feature vector ( pFor the preset feature dimensions, q (Preset feature length); the decoder reconstructs the input data through deconvolution layers, and the loss function is defined as the sum of the three-channel mean square errors of the input and the reconstructed output:

[0069]

[0070] in, x k This refers to the input sample. x k recon This refers to reconstructing samples; intermediate features are mapped to standardized samples via a fully connected layer. p*q The feature vectors are used as input features for the capacity prediction model below.

[0071] Step Four: As Figure 2 As shown, a data-driven model using GCN-BiLSTM is constructed to estimate the remaining charge capacity of individual battery cells. Specifically, this includes:

[0072] During the data construction process, the intermediate features are first divided into equal parts of a fixed length. w Each time segment is used to calculate the mean value of a node feature vector, and these nodes together form the vertex set of the graph. Then, the cosine similarity between the features of the nodes is calculated to establish the connection relationship of the edges. When the similarity between two nodes exceeds a preset threshold, a directed edge is established between the two nodes. All node pairs that meet the conditions will be connected by edges, and finally a directed graph structure based on the similarity of time series data is formed. The nodes retain the local statistical features of the original time series data, while the edges reflect the correlation strength between different time segments. This model employs a hybrid architecture of Graph Convolutional Neural Network (GCN) and Bidirectional Long Short-Term Memory (BiLSTM). During construction, a two-layer GCN (GCNConv) is first used to extract features and perform message passing on the input graph data. Each GCN layer is followed by a ReLU activation function to enhance non-linear expressiveness. Subsequently, graph-level feature pooling is performed using global_mean_pool to aggregate the features of all nodes into a unified graph representation. Next, the pooled features are input into a BiLSTM layer for temporal modeling to capture long-term dependencies in the sequence. Finally, a fully connected layer maps the hidden states of the BiLSTM to the prediction target space, thus completing the end-to-end mapping process from graph structure features to the final predicted value. The entire network model is trained to optimize the loss function. Mean Absolute Error (MAE) is chosen as the loss function to quantify the difference between the predicted and true values. The formula is as follows:

[0073]

[0074] in, y i The actual value is shown below, while the model predicts the value. m The sample size is specified. During model training, the entire process is built on the PyTorch framework. The model structure employs a hybrid architecture of GCN and BiLSTM. The GCN layer contains two layers of graph convolutional networks, each followed by a ReLU activation function, while the BiLSTM layers are used for sequence modeling. During training, the Adam optimizer is used for parameter optimization, with an initial learning rate of 0.005 and a total of 1000 training epochs. Model training uses batch processing with a batch size of 512, utilizing DataLoader for batch loading and random shuffling of data. In each training epoch, the model sequentially performs forward propagation to calculate predicted values, uses L1Loss to calculate the loss, and then updates the model parameters through backpropagation.

[0075] Step 5: Based on the trained model, calculate the remaining rechargeable capacity (RCC) for each unit in each cycle:

[0076] First, according to the method of S21, input samples are extracted from the charging data of each charging cycle, and then the samples are input into the trained model to estimate the value of each individual in the first charging cycle. j The remaining charging capacity after one charge cycle to the upper cutoff voltage. Q j =[ Q 1j ，Q 2j ，…，Q ij ，Q Nj ],in N Indicates that the battery pack has a total of N Individual battery cells, Q ij Indicates the first i The monomer in the first j The remaining charging capacity after reaching the upper cutoff voltage in one cycle.

[0077] Then, due to different battery charging strategies, the current switching points are not entirely the same. This means the starting point for calculating the remaining rechargeable capacity is inconsistent. To ensure a relatively stable RCC value in each calculation, we need to... Q j Normalization is performed to obtain the RCC values ​​of cells that cannot reach the upper cutoff voltage, thereby eliminating system bias and improving the stability of the calculation results. The specific calculation formula is as follows:

[0078] in,RCC ij In the j-th loop, the first... i The remaining rechargeable capacity of each cell; min ( Q j ) is the first j The minimum remaining charge capacity of all cells in each cycle. The estimated RCC of all cells in the j-th cycle is obtained. RCC j = [ RCC 1j ， RCC 2j ,..., RCC Nj ]:

[0079]

[0080] Step Six: The RCC (Recovery Capacity) of a lithium battery can effectively reflect inconsistencies between individual cells. Generally, cells with smaller RCCs have higher SOCs and smaller capacities, while cells with larger RCCs have lower SOCs and larger capacities. Based on the size, distribution, and variations of the RCCs of each cell in the battery pack, we can diagnose overall battery pack inconsistencies, individual cell SOC inconsistencies, and self-discharge faults. Figure 3 As shown:

[0081] First, the overall inconsistency of the battery pack is measured. To quantify this inconsistency, the range of RCC values ​​is used as the metric. The range is defined as:

[0082]

[0083] Where: max( RCC j ): No. j The maximum value of the individual RCC in each cycle; min( RCC j ): No. j The minimum RCC of a single unit in each cycle. Range RCC This range directly reflects the range of RCC differences among individual cells within the battery pack and can be used to characterize the consistency of the battery pack. The larger the range, the more severe the inconsistency among the individual cells within the battery pack.

[0084] Next, regarding the inconsistency of SOC in individual cells, the difference between the RCC value of a single cell and the median RCC can be calculated. The median is chosen for its noise resistance and accurate reflection of the central tendency of the data. The formula is as follows:

[0085]

[0086] in: D ij : No. j In the loop, the first... i The difference between the RCC of each individual and the median; Median ( RCC j ): No. j The median of the RCC set in each cycle. Individual difference. D ij It characterizes the deviation of its RCC from the concentration level of the battery pack, thus reflecting the inconsistency of the SOC of the cell.

[0087] Secondly, when a single cell in the battery pack experiences a self-discharge fault, its RCC value will exhibit an abnormal decrease, and the slope of its RCC value will differ significantly from that of a normal cell. Based on this principle, we can use the sliding window local consistency analysis method to calculate the RCC slope of each cell in the battery pack, thereby identifying the self-discharging cell. The sliding window local consistency analysis method is a highly efficient anomaly detection method, particularly suitable for time-series data with high noise or significant fluctuations. Through this method, we can further diagnose battery self-discharge faults.

[0088] First, define the sliding window: based on the sampling frequency of the battery pack cycling data and the analysis requirements, set a reasonable window length. W The window size can be set to 5 or 10 cycles, depending on the volatility of the data. An appropriate sliding step size should also be selected. Δ (e.g., 1 or 2) Ensure the window covers all data points. The range of the sliding window is... T k =[ j , j + W -1], the formula for calculating the rate of change of RCC is:

[0089]

[0090] △ RCC i,k It is a matrix representing the RCC of each battery cell in the window. T k Rate of change within.

[0091] Calculation within the time window T k The local mean within the time window is used to calculate the local deviation of each individual within this time window, thereby measuring the consistency of change of each individual within this time window. The specific formula is as follows:

[0092]

[0093] Where: indicates within the time window T K The local mean of each time window within the period, while S i,k It is the local deviation, which represents the difference between the individual change and the local average change.

[0094] Finally, the RCC range for each cycle can be calculated based on historical data. RCC The difference between the RCC of a single monomer and the median D ij Use the 95th percentile to set the threshold. thrshold1 , thrshold2 This is used to assess the overall inconsistency of the battery pack and the inconsistency of the state of charge (SOC) of individual battery cells. It also involves calculating local deviations based on historical data. S i,k 95th percentile threshold setting thrshold3 and multiple consecutive time windows exceeding the threshold setting thrshold4 To locate and diagnose the faults of self-discharging cells.

Claims

1. A method for diagnosing self-discharge of abnormal individual cells in a battery pack, the method being applicable to multi-stage variable charging conditions, the method estimating the remaining rechargeable capacity of the RCC based on short-time charging data and a GCN-BiLSTM graph convolutional network-bidirectional long short-term memory network algorithm, and then performing self-discharge diagnosis of abnormal individual cells in the battery pack; characterized in that, Includes the following steps: S1. The battery pack operation data during the charging phase is collected in real time by the battery management system and preprocessed. The data includes total voltage, current, extreme values ​​of individual cell voltage, extreme values ​​of temperature, and individual cell voltage sequence. S2. Extract charging data at a fixed length from the current switching point and calculate the remaining charging capacity of the individual cell when the upper cutoff voltage is reached. Q As training samples, the sample data is normalized. S3. Construct a multi-condition adaptive feature extraction model with an encoder-decoder architecture. By reconstructing the input, the input data under different conditions are mapped to the same latent space to extract intermediate features. S4. Construct a hybrid model based on graph convolutional network and bidirectional long short-term memory network. Extract graph structure features through two layers of GCNConv, and input them into BiLSTM for time series modeling after global pooling to establish the mapping relationship between intermediate features and remaining charging capacity. S5. Input the real-time charging data into the trained model, calculate the remaining rechargeable capacity RCC value of each cell, and obtain the standardized RCC index through range normalization. S6. Based on the range distribution, median deviation and sliding window local consistency analysis of RCC values, the overall inconsistency assessment of battery packs, and the diagnosis and separation of SOC inconsistency faults and self-discharge faults are realized. The data acquisition and preprocessing in step S1 includes: recording the total voltage, total current, extreme values ​​of individual cell voltages, extreme values ​​of temperature, and all individual cell voltage sequences at a preset sampling frequency; performing multi-level cleaning on the raw data; using sliding window mean filtering to eliminate high-frequency noise; identifying and removing abnormal voltage jump points based on the box plot method; using linear interpolation to fill in missing data segments; and constructing a feature matrix X∈R^{N×L×3} with timestamps, where... N For the number of individual units, L For the time step, the 3D features include voltage, current, and temperature; The sample construction method in step S2 includes: Define current switching event detection conditions: when the rate of change of charging current exceeds a preset threshold. ΔI_threshold At that time, data interception is triggered; individual cells in the battery pack that have reached the upper cutoff voltage are selected, and the voltage-current-temperature three-dimensional sequence of n sampling points is symmetrically intercepted with the switching point as the center, as the input sample: For a single cell that has reached the upper cutoff voltage, the remaining charging capacity is calculated using the ampere integration method, where the time range is […]. T0, T1 The definition is as follows: T0 The start time of the remaining charging segment of a single battery cell; T1 The actual end time of charging a single battery cell is the point at which charging reaches the charging cutoff voltage. Then, the remaining charging capacity of that single cell is calculated using the Ampere integral formula. Q : The time range is [ T0 , T1 Within [the specified range], extract the current data sequence of the individual cell. I(t) ; The encoder-decoder architecture in step S3 includes: constructing an encoder-decoder network, with the input being the voltage of the charging data samples. U Current I ,temperature T A three-channel temporal sequence; the encoder extracts features through three one-dimensional convolutional layers and compresses them into the latent space, with an output dimension of [missing value]. p * q The intermediate feature vector, where p For the preset feature dimensions, q The preset feature length is used; the decoder reconstructs the input data through deconvolution layers; the loss function is defined as the sum of the three-channel mean square errors of the input and the reconstructed output: in, x k' This refers to the input sample. x k' recon This refers to reconstructing samples; intermediate features are mapped to standardized samples via a fully connected layer. p*q The feature vectors are used as input data for the capacity prediction model below; The hybrid model construction in step S4 includes: a graph structure construction module: dividing the input intermediate features into equal parts of a fixed length. w Each time segment is used to calculate a node feature vector by averaging the features. These nodes together form the vertex set of the graph. Then, the cosine similarity of the features between nodes is used to establish edge connections. When the similarity between two nodes exceeds a preset threshold, a directed edge is established between them. All node pairs meeting the condition are connected by edges, ultimately forming a directed graph structure based on the similarity of time-series data. Nodes retain the local statistical features of the original time-series data, while edges reflect the correlation strength between different time segments. The model building module uses a hybrid architecture of GCN and BiLSTM. First, a two-layer GCNConv with ReLU activation is used for graph feature extraction and message passing. After global mean pooling to aggregate graph-level features, the data is input into BiLSTM to capture temporal dependencies. Finally, a fully connected layer is used to achieve end-to-end target mapping prediction. The entire network model is trained to optimize the loss function: the mean absolute error (MAE) is chosen as the loss function to quantify the difference between the predicted and true values. The formula is as follows: in, y i For the true value, y i p These are the model's predicted values. m The number of samples; The RCC calculation in step S5 includes: processing the original output of the model... Q The values ​​are dynamically baseline corrected; input samples are extracted from the charging data of each charging cycle; and input samples are symmetrically extracted from n sampling points centered on the switching point. Sample ij : indicates the first i The first monomer j The input samples are processed in a loop; then the samples are fed into the trained model to estimate each individual in the 1st cycle. j The remaining charging capacity after one charge cycle to the upper cutoff voltage. Q j =[ Q 1j ，Q 2j ，…，Q ij ，Q Nj ],in N Indicates that the battery pack has a total of N Individual battery cells, Q ij Indicates the first i The monomer in the first j The remaining charging capacity after each cycle reaches the upper cutoff voltage; for Q j Offset adjustment is performed to obtain the RCC value of the individual cells that cannot reach the upper cutoff voltage, in order to eliminate system bias and improve the stability of the calculation results; the specific calculation formula is as follows: in, RCC ij In the j-th loop, the first... i The remaining rechargeable capacity of each cell; min ( Q j ) is the first j The minimum remaining charge capacity of all cells in each cycle is obtained; the estimated RCC of all cells in the j-th cycle is obtained. RCC j = [ RCC 1j ， RCC 2j , ..., RCC Nj ]; The diagnostic mechanism in step S6 includes: firstly, measuring the overall inconsistency of the battery pack, and calculating the range of RCC values ​​as a measurement indicator: Where, max( RCC j ): No. j The maximum value of the individual RCC in each cycle; min( RCC j ): No. j The minimum RCC of a single cell in each cycle, and the range. RCC This directly reflects the range of RCC differences among individual cells within the battery pack, and can be used to characterize the consistency of the battery pack. The larger the range, the more severe the inconsistency among the individual cells within the battery pack. Next, the SOC inconsistency of individual cells is evaluated by calculating the difference between the individual cell RCC value and the median RCC. D ij : in, D ij : No. j In the loop, the first... i The difference between the RCC of each individual and the median, Median( RCC j ): No. j Median of the RCC set in each cycle; Individual difference D ij The deviation of the RCC from the concentration level of the battery pack was characterized, thus reflecting the SOC inconsistency of the individual cell. Finally, the sliding window local consistency analysis method was used to calculate the RCC slope of each cell in the battery pack, thereby determining the self-discharge cells. First, the sliding window was defined: the window length was set according to the sampling frequency of the battery pack cycle data and the analysis requirements. W The window length W Choose between 5 or 10 cycles based on the volatility of the data, and select the sliding step size. Δ The sliding step size Δ Choose option 1 or 2 to ensure the window covers all data points; the range of the sliding window is... T k =[ j , j+W -1], the formula for calculating the rate of change of RCC is: Among them, △ RCC i,k It is a matrix representing the RCC of each battery cell in the window. T k Rate of change within a time window; calculated within a time window T k The local mean within the time window is used to calculate the local deviation of each individual within this time window, thereby measuring the consistency of change of each individual within this time window. The specific formula is as follows: Wherein, ΔRCC k ave Indicates within the time window T K The local mean of each time window within the period; in, S i,k It is the local deviation, representing the difference between the individual change and the local average change; finally, the RCC range is calculated for each cycle based on historical data. RCC The difference between the RCC of a single monomer and the median D ij Use the 95th percentile to set the threshold. threshold1 , threshold2 To assess the overall inconsistency of the battery pack and the inconsistency of the state of charge (SOC) of individual battery cells, and to calculate the local deviation based on historical data. S i,k 95th percentile threshold setting threshold3 and continuous z The time window exceeds the threshold setting threshold4 To locate and diagnose the faults of self-discharging cells.

Images